CS 337 -- Intro to Semantic Information Processing -- L. Birnbaum






Natural Language Understanding is the art & science of getting computers to UNDERSTAND natural



What does computer science contribute to the study of language?


    A stress on PROCESS: Computer science provides a set of concepts

    for describing HOW it is accomplished.


    But more important, a stress on FUNCTION: What is language FOR?

    The goals which a language processor must satisfy stem, primarily,

    from the goals of the task which it must accomplish.


The purpose of language is to enable communication of ideas, feelings,

needs.  The speaker has something in his mind which he wishes the

hearer to know.


The basic model:


    IDEAS --generation--> LANGUAGE --understanding--> IDEAS


Ideas include, among other things, facts and opinions about the world:



REPRESENTATION and PROCESS are the issues:


    We need a way to represent the ideas.


    We need a way to encode (generate) and decode (understand, parse)

    ideas in language.


These two facets are interdependent:


    Representations impose requirements on processes: The processes

    must take certain representations as input and produce the

    appropriate representations as output.


    Processes impose requirements on representations: The

    representations must, as much as possible, facilitate the processes

    that operate on them.


This is just common sense, but compare linguistics and psychology

(gently): The lack of teleological reasoning, and hence of

functionally adequate theories, is striking.


Back to the model: People don't have empty heads -- the problem is,

computers do.


    The ideas stem from the speaker's thought processes.


    They engender thought processes in the hearer -- they INTERACT

    with other ideas.  The hearer must:


      Remember them.


      Relate them to other things he knows.


      Draw appropriate conclusions.


This is of course true regardless of whether or not the communication

is linguistic, or indeed whether the task is communication at all.


    The same issues arise in perception, planning, problem-solving --

    all cognitive functions.


    The problem of understanding language is the problem of

    understanding IDEAS.


Understanding a text or an utterance involves understanding the events

described in that text or utterance -- that is, being able to EXPLAIN

what happened and why -- and understanding why the speaker or writer

spoke or wrote what he did as he did -- that is, being able to EXPLAIN

the goals which the text or utterance serves.


We can see how well people understand by asking them questions:


    When John started his freshman year at Northwestern, his parents

    told him they would buy him a BMW if he got good grades.  John

    studied very hard that year.


    Question:  Why did John study hard?


Understanding requires making INFERENCES --


    -- About what the text describes:


      John went to a store.  He picked up some toothpaste and went

        to the check-out counter.  He paid the cashier and left.


      Question:  What did John buy?


      Being able to answer this question depends on being able to

      understand John's actions -- why he is doing what he is doing.


    -- About why the speaker said it:


      "Do you know the time?"


      Question:  Why does the speaker want to know if you know?


      "I'm hungry."


      Question:  Why does the speaker want you to know?


The need for inference is underscored by linguistic problems.

Language is:




      word sense ambiguity -- e.g., of the word "sense"


      structural ambiguity -- "I saw the Grand Canyon flying to New



    vague: "I got a new TV at Marshall Field's."


    elliptic: "John thinks vanilla."


            "What flavor ice-cream does Mary like?" (McCawley)


    metaphorical: "There's a cancer in the White House." (John Dean)


We are able to communicate ideas using a system with these

characteristics because of SHARED KNOWLEDGE, and our ability to draw

inferences from what we are told and what we already know.


The sheer amount of this knowledge is staggering.  Big issues:


    How can all of this knowledge be represented in a computer?


    How can it be brought to bear at the appropriate time and in the

    appropriate way?


Bar-Hillel's example: Lexical disambiguation and knowledge of size,

space, and function.


    The box is in the pen.


    (Compare with: The pen is in the box.)


Winograd's example: Pronoun reference resolution and political



    The city council refused the demonstrators a permit because they

    feared violence.


    The city council refused the demonstrators a permit because they

    advocated violence.


Another variant:


    Chicago, IL -- The city council refused the demonstrators a permit

    because they were communists.


    Warsaw, Poland -- The city council refused the demonstrators a

    permit because they were communists.


Learning to read naively is the first step to understanding the

problems of Natural Language Understanding.


NY Times example #1:


    An Arms Obstacle Falls: Moscow Puts Aside 'Star Wars' Demand,

    Removing a Bar to Strategic-Weapons Pact.


    Word-sense ambiguity of "arms", "bar", and "falls"


    vagueness/metaphor of "falls", "remove"


    ellision of "arms obstacle" = obstacle to arms treaty, as opposed

    to (say) an obstacle to arms construction


    structural ambiguity of "to Strategic-Weapons Pact" -- is this

    attached to "bar" or "remove" (as in, "National Guard Removes 100

    People to Safety in Charleston")?


    metaphor of "falls", "bar", and "remove" -- the goal of the treaty

    is at the end of a road, and something has been blocking progress,

    and that something is now gone


NY Times example #2:


    In "Annie Hall," [Woody Allen] observed that Los Angeles's great

    cultural contribution was the right turn on the red light.


    Looking at 3 levels of difficulty:


    Word-sense ambiguity of "observed" -- communicating an observation

    vs. just making one:


      In "Annie Hall," millions of viewers observed that Woody Allen

      was as neurotic as ever.


    Ellipsis of "right turn on red light:" What does this phrase refer

    to, and what do you need to know to understand it?


    Ironic implications of the whole statement: What does this

    sentence really mean, and what do you need to know to recognize the



NY Times example #3:


    U.S. Said to Soften Stand on Missiles at Geneva Parley.


    The ambiguity of "soften."


    The metaphorical nature of "soften stand": How do we interpret

    this as "reduce demands"?


    Ambiguity of relation denoted by "on": Understanding the metaphor

    lets us realize that this means "concerning," as opposed to

    "physically connected to" as in "U.S. said to improve launch stand

    on mobile missiles."


    Structural ambiguity of "at Geneva parley": it modifies "soften

    stand," not "missiles."  Compare:  "U.S. said to soften stand on

    children at state dinners."


    Which talks are these, anyway?


NY Times example #4:


    The United States and the other Western nations with forces in

    Lebanon have turned down an urgent Lebanese Government request that

    they enter the disputed Shuf area around Beirut and try to halt the

    growing civil strife there, according to Secretary of State George

    P. Shultz and other Administration officials.


    "Western" = Western world, not Western U.S. or genre of film


    "forces" = military forces, not magnetic


    "turned down" = refused or rejected, not "adjust lower" as in

    "turn down the heat;" nor as in "turn down the bedcovers"


    "disputed" = conflicting claims of control; related to, but not

    the same as, disputing a fact


    "growing" = increasing, but not as in "growing a garden"


    "civil" does not mean polite in this case


    "in Lebanon" applies to forces, not nations; compare "men with

    guns in Lebanon"


In these examples, we see how inference based on context and

background knowledge help to solve low-level linguistic problems.


    This is how the need for plausible inference and world knowledge

    in language understanding was first discovered and justified.



    Inference and memory processing have their own purpose and dynamic.


    We ALWAYS try to explain inputs -- linguistic or perceptual -- and

    relate them to what we know.  Linguistic problems are solved as a

    result of pursuing this goal.





Analyze a paragraph from a magazine or a book, and try to uncover as

many of the above-illustrated problems as you can -- ambiguous words,

ambiguous constructions, implicit content, vagueness, metaphor, irony,

and others that you will discover.  For each problem you discover, try

to determine what facts you knew that helped you to resolve the

problem and understand the paragraph.  These are the facts a computer

program would need as well.






1.  The very fact that we can distinguish the ambiguous senses of the

words and utterances shows that we have some way to represent the

distinct meanings.  In other words, our internal representation of an

utterance is not ambiguous, at least, not to the same degree that

English is.


2.  Deriving a sufficiently unambiguous meaning representation of an

utterance requires, to an astonishingly high degree, the use of

inference applied to the text in conjunction contextual knowledge and

background world knowledge.


3.  The inferences reveal implicit content in utterances, which needs

to be represented as well.


4.  Thus, we must be concerned not only with how to represent the

meaning of the utterance in a sufficiently disambiguated and explicit

way, we must worry about how to represent the knowledge that is being

used to construct that representation.  A common representation for

both text meanings and world knowledge would be desirable.


5.  We must be concerned with how this knowledge is organized so as to

enable access and application during understanding.